Useful knowledge arrives, but gets distorted.

The knowledge got there, but it changed on the way. It was softened on its way up, squeezed into a summary, pulled toward the first answer in the room, or made more convincing by fluent AI output. Because the information still looks credible, the distortion can be hard to notice.

Compare what arrived with something independent of it.

• Put the original next to the version that reached the decision: the front-line account next to the report upward, or the analyst's finding next to the summary two levels up.
• Check a sample of claims against their sources. For AI output, check one claim in each of five recent pieces that went out.
• Find out when people formed their views: before or after they saw the AI's answer or heard the senior person's view. Count how many of the options on the table really differ.
• Look at the review steps: how often they approve, how long a review takes compared with a careful check, and whether they have ever stopped anything.

Fixes for corruption protect independent views and add checks that actually catch things: views written down privately before discussion, a named checker who isn't the author, a check against a source that fails differently from the AI, and enough review capacity for the volume of work.